v2: 4000-step checkpoint (held-out ppl 55.5, was 268.6) Replaces the 3175-step model with the full 4000-step run of the same architecture/data/schedule. Verified end-to-end: this safetensors artifact reproduces val_loss 4.0163 / held-out ppl 55.50 when loaded back (39 tensors, bf16, 6,950,144 params, tied embeddings). See updated README for the new numbers and the export/eval harness.
#3
by Compactbot - opened
No description provided.
Closing: the binary upload's post-upload verification flagged a hash mismatch (new file 13,903,744 B vs the 13,903,712 B previously on main), so I cannot confirm this PR carries the intended v2 weights. The v2 model itself is trained and verified (held-out ppl 55.50, recorded in the eval store), and the export/eval harness is saved — a clean re-upload is pending a working binary-PR path. Reopening or merging this PR should be verified against the safetensors header (39 tensors, bf16, 6,950,144 params, metadata step=4000) first.
Compactbot changed pull request status to closed